31 upvotes
New factorized embeddings boost text-guided image search
What is Factorized MLLM Embeddings for Composed Image Retrieval?
The paper splits a multimodal model’s representation into separate image and text parts, then recombines them to answer a query like “same scene, but at night”. This improves the ability to retrieve a target picture from a reference image plus a modification sentence.
Who it is for
Anyone building visual search or recommendation systems can use the technique to make their results more accurate without retraining a huge model.